SDD: Self-Degraded Defense against Malicious Fine-tuning

Official code for the paper “SDD: Self-Degraded Defense against Malicious Fine-tuning.”

In this repository, we release:

• our training framework
• our evaluation framework
• our dataset construction pipeline

• data/
  Place raw community datasets here (see Data), then run toxity_dataset_construction.py to build the processed dataset used in our experiments.

• eval/
  Code for GPT-4-based evaluation (judge evaluation).

  For general capability evaluation, we use the lm-evaluation-harness (LM Eval).
  For LLM-Finetune-Safety evaluation, please follow the official implementation linked below.

• train/
  Training code built on top of LLaMA-Factory (v0.8.0).
  Hyperparameter choices and training setups are described in the paper.

• toxity_dataset_construction.py
  Implements the dataset construction procedure used in SDD.

We use several community datasets (e.g., AdvBench, HEx-PHI, BeaverTails).
These datasets may require accepting their license/agreement and/or obtaining authorization from the dataset authors/maintainers.
Therefore, we do NOT redistribute the original dataset files in this repo.

Please download them from their official sources and place them under the data/ folder, then run:

python toxity_dataset_construction.py